PeopleofSuperintelligence

Builder

Jiashi Li

Jiashi Li works with the DeepSeek-AI team in Hangzhou on the GPU kernels and computing infrastructure behind DeepSeek's large language models. In 2024, Li coauthored DeepSeekMoE, DeepSeek-V2 with its Multi-head Latent Attention, the Fire-Flyer AI-HPC paper on a cluster of 10,000 A100 GPUs, and the DeepSeek-V3 technical report. They were first author of FlashMLA, DeepSeek's library of attention kernels, in 2025, and coauthored the DeepEP and DeepGEMM libraries, an ISCA 2025 study of the hardware behind DeepSeek-V3, and the DeepSeek-R1 paper in Nature. Their 2026 papers include DSpark, a speculative decoding framework for DeepSeek-V4 serving, and the DeepSeek-V4.1-Flash model, whose inference on NVIDIA GPUs and Huawei Ascend NPUs runs on FlashMLA.

Now at
DeepSeek

Timeline

  1. January 2024

    The DeepSeekMoE architecture

    In January 2024, Li coauthors DeepSeekMoE, a mixture-of-experts architecture that splits experts into finer segments and isolates some of them as shared experts. Scaled to 16 billion parameters, the model matches LLaMA2 7B with about 40% of the computation.

  2. May 2024

    Multi-head latent attention in DeepSeek-V2

    DeepSeek-V2, which they coauthor in May 2024, is a mixture-of-experts model with 236 billion total parameters and 21 billion activated for each token. Its Multi-head Latent Attention (MLA) compresses the key-value cache into a latent vector for efficient inference.

  3. August 2024

    The Fire-Flyer 2 cluster

    Li coauthors the Fire-Flyer AI-HPC paper in August 2024, a hardware and software co-design whose Fire-Flyer 2 cluster of 10,000 PCIe A100 GPUs approaches DGX-A100 performance at half the cost and with 40% less energy. The paper is accepted for presentation at the SC'24 high-performance computing conference in 2024.

  4. December 2024

    The DeepSeek-V3 technical report

    In December 2024, they coauthor the DeepSeek-V3 technical report on a mixture-of-experts model with 671 billion total parameters and 37 billion activated for each token. Its full training takes 2.788 million H800 GPU hours.

  5. 2025

    FlashMLA attention kernels

    In 2025, Li is first author of FlashMLA, DeepSeek's library of optimized kernels for Multi-head Latent Attention. They also coauthor DeepEP, a communication library with expert-parallel GPU kernels for mixture-of-experts models, and DeepGEMM, a tensor core kernel library for large language models.

  6. May 2025

    A DeepSeek-V3 hardware study

    A May 2025 paper they coauthor analyzes how DeepSeek-V3, trained on 2,048 NVIDIA H800 GPUs, pairs model design with hardware through Multi-head Latent Attention, FP8 mixed-precision training and a multi-plane network topology. The paper appears in the Industry Track of ISCA 2025, the International Symposium on Computer Architecture.

  7. September 2025

    DeepSeek-R1 in Nature

    As part of the DeepSeek-AI team in Hangzhou, Li coauthors the DeepSeek-R1 paper that Nature publishes in September 2025. It shows that pure reinforcement learning can develop reasoning in large language models without human-labelled reasoning trajectories.

  8. As of October 2026

    DSpark and DeepSeek-V4.1-Flash

    In July 2026, Li coauthors DSpark, a speculative decoding framework that speeds per-user generation by 60 to 85 percent in the DeepSeek-V4 serving system. They appear among the authors of DeepSeek-V4.1-Flash in September 2026, a multimodal model with 552 billion backbone parameters and contexts of up to one million tokens. FlashMLA powers DeepSeek-V4.1 inference on NVIDIA GPUs and Huawei Ascend NPUs.

  1. The median h-index is 48 at universities and 14 at frontier AI labsIncludes DeepSeek. The median h-index is 48 at universities and 14 at frontier AI labs. We compare publishing careers, recent citations, dual appointments and job moves across 2,420 university researchers and 308 frontier-lab researchers.
  2. Companies supply the most prior-work hires at 7 of the 9 frontier AI labsIncludes DeepSeek. Our estimates of the organizations where 550 people worked directly before joining 9 frontier AI labs across 620 career transitions, comparing the main feeder organizations per lab.
  3. 5,000 frontier users would be 38.2% of the core labs' 13,100 reported staffIncludes DeepSeek. 5,000 frontier users would be 38.2% of the core labs' 13,100 reported staff. We test the estimate against 3,913 named model contributors and dated all-staff records across Google DeepMind, OpenAI and Anthropic, plus six additional frontier-grade labs.
  4. Google alumni have founded or led 133 other organizations, more than any other AI labIncludes DeepSeek. We estimate that alumni of 17 major AI labs went on to found or lead 350 other organizations, including 226 they founded, and we trace 472 career moves directly between 23 labs since 2010.

Sources and credits

Photo: We have no published photo of Jiashi Li yet. Send us a free-licensed photo.

  1. DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models arxiv.org. Checked 20 October 2026.
  2. DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model arxiv.org. Checked 20 October 2026.
  3. Fire-Flyer AI-HPC: A Cost-Effective Software-Hardware Co-Design for Deep Learning arxiv.org. Checked 20 October 2026.
  4. DeepSeek-V3 Technical Report arxiv.org. Checked 20 October 2026.
  5. GitHub - deepseek-ai/FlashMLA: FlashMLA: Efficient Multi-head Latent Attention Kernels github.com. Checked 20 October 2026.
  6. GitHub - deepseek-ai/DeepEP: DeepEP: an efficient expert-parallel communication library github.com. Checked 20 October 2026.
  7. GitHub - deepseek-ai/DeepGEMM: DeepGEMM: clean and efficient BLAS kernel library on GPU github.com. Checked 20 October 2026.
  8. Insights into DeepSeek-V3: Scaling Challenges and Reflections on Hardware for AI Architectures arxiv.org. Checked 20 October 2026.
  9. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning nature.com. Checked 20 October 2026.
  10. DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation arxiv.org. Checked 20 October 2026.
  11. DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression arxiv.org. Checked 20 October 2026.